What you are actually paying for
- Model usage per request, scaled by how much context is supplied.
- Each step in a multi-action chain, since each is its own model call.
- Retries after a failure or an ambiguous instruction.
- Any external service the action touches, such as messaging or documents.
- Human review time on confirmation gates and escalations.
Why per-seat pricing fits badly
Seat pricing assumes cost follows people. Operational AI cost follows volume: one dispatcher issuing fifty instructions a day costs more than ten people who each ask one question.
Usage-based models track that honestly, which is why the important question is not the rate but the ceiling.
How to keep spend bounded
- Set a hard cap per organization and per period, not just an alert.
- Bound context: send the records the action needs, not the whole history.
- Cap loop iterations so a confused chain cannot spiral.
- Reserve budget before a chain runs and release the unused portion.
- Report consumption per user so outliers are visible early.
The comparison that matters
The useful comparison is not AI cost against zero. It is AI cost against the labour it displaces plus the revenue currently lost to work that simply never gets done — unchased estimates, uncollected invoices, unreturned calls.
That comparison is specific to your volumes. Anyone quoting a universal return figure is guessing.